A LIGHTWEIGHT ATTENTION-ENHANCED CNN FOR CERVICAL CELL CLASSIFICATION

Abstract

Accurate and early identification of abnormal cervical cells is essential for reducing cervical cancer mortality; however, current deep learning solutions often rely on computationally heavy archi-tectures that limit clinical deployment. To address this challenge, we propose a lightweight attention-enhanced convolutional neural network (AECNN) that integrates depthwise separable layers with spa-tial and channel squeeze-and-excitation attention. The model is designed to emphasize nucleus-specific morphological cues while maintaining low computational cost. In contrast to recent attention-guided frameworks such as AEDA [11] and LAHN [12], the proposed architecture introduces hybrid class-balancing through weighted cross-entropy and focal loss to mitigate false negatives and improve sensi-tivity in minority classes. Experiments conducted using SIPaKMeD for training, Herlev for validation, and CRIC for independent testing demonstrate significant improvements in diagnostic performance. The proposed AECNN achieved a sensitivity of 92.14%, specificity of 94.63%, accuracy of 93.87%, and an AUC of 0.958, outperforming AEDA and LAHN by margins of 3.2–5.6% depending on the metric. These results suggest that the proposed lightweight design offers a clinically viable balance between high discriminatory capability and computational feasibility, supporting the potential integration of the method into real-time screening applications.